Games have long been a great way to benchmark how sophisticated computers are becoming. From Deep Blue achieving the first computer victory over a reigning world champion in 1996, to AlphaGo's famous Go win last year, slowly but surely computers are beating us at every game.

It looks like the next to fall will be poker. AI system Libratus is slowly building a lead over four top humans in a 120,000 hand challenge. Poker is a step-up over the previous games conquered by AI, with the challenges of incomplete information, bluffing and other tactics making it more difficult for a computer to outplay humans. A whopping 15 million hours of computing has been used to build Libratus, that's no mean investment.

I'm interested though, in how these AI systems are being trained up on a pure diet of the game they're trying to beat. Libratus just knows poker, and I'd bet would be rubbish at Monopoly. Meanwhile the top human players may spend most of their time on poker, but have grown up and learnt by playing everything from Connect Four to Grandmother's Footsteps. Maybe the next challenge for AI shouldn't be beating humans in another complicate game, but building a general game playing engine that can learn and adapt to any new game, without millions of hours of training behind it.